hosted opensource + the OmniRoute gateway

Same opensource models. Hosted. One endpoint.

Not every rig can run a 70B model. OmniRoute is an MIT-licensed gateway that stacks the free tiers of 290+ providers into one local OpenAI-compatible endpoint. Point vinos-ai and your routines at it and you get frontier opensource models — Kimi K2/K3, DeepSeek, GLM, Llama, Qwen — without a single API key on day one. ~1.53B free tokens per month, auto-fallback when one provider throttles, works out of the box. Complement to /models/ (local Ollama) — use both.

OmniRoute upstream, re-audited every two weeks. Free-tier math lives at docs/reference/FREE_TIERS.md — a provider ends a free tier and the number drops; a new one lands and it climbs. Numbers below reflect the current catalog.

Zero-config quickstart

Two commands. First one installs OmniRoute; second points vinos-ai at it. That's it.

npm i -g omniroute # boots on http://localhost:20128 export VINOS_AI_BASE_URL=http://localhost:20128/v1 export VINOS_AI_MODEL=auto # OmniRoute picks the cheapest working provider

vinos-ai chat “hello” # answers via a free-tier provider — no API key

Pre-wired keyless providers (OpenCode Free, Felo) mean auto works before you sign up anywhere. Add API keys later for Kimi, DeepSeek, Groq, etc. — OmniRoute layers them in and prefers the cheapest one that's still under quota. Endpoint reference: OmniRoute quick start.

The 4-tier auto-fallback

Every call goes through OmniRoute's smart router. If the top-tier provider is out of quota, or the request fails, or the response confidence is low — it slides to the next tier silently, in milliseconds. No retries in your code, no downtime.

Tier 1 — Subscription Claude Code plan, Copilot plan, Codex plan (if you already pay)

Tier 2 — API key DeepSeek, Groq, xAI, Anthropic direct — the ones you brought your own key for

Tier 3 — Cheap paid GLM $0.50/M, MiniMax $0.20/M — routed when you're willing to spend pennies

Tier 4 — Free tier 90+ free-tier providers, ~1.53B tokens/mo aggregate — the default fallback

Configure vinOS routines

Same routine TOML as local routing — the router still runs local first for the 80% case. The premium_model now points at OmniRoute instead of Anthropic direct, so escalations go through the gateway's free/cheap tiers before spending real money.

[agent]
route          = "auto"                      # anthropic | ollama | omniroute | auto
local_model    = "qwen2.5:7b"                # local Ollama for the 80% case
premium_model  = "auto/smart"                # OmniRoute meta-model — picks best free provider
[agent.omniroute]
base_url       = “http://localhost:20128/v1”
api_key        = “sk-any-string”             # OmniRoute ignores it locally
[agent.escalation]
on_low_confidence   = true
on_reasoning_task   = true
on_context_overflow = true
max_escalations_per_run = 3

Model IDs to try: auto (balanced), auto/coding (code-quality-first), auto/fast (lowest latency), auto/cheap (cheapest first), auto/smart (quality-first + exploration). OmniRoute maps each of these to a live combo across whichever providers are currently under quota.

When local, when hosted?

Use local (/models/) Offline / airgapped work · sensitive data · always-on daemons · when you own the GPU · sub-second latency

Use hosted (this page) Frontier models your box can't run (Kimi K3, Claude Sonnet, GPT-4o) · burst workloads · rigs without a GPU · free-tier stacking on principle

Use both (recommended)
route = "auto" — local for the 80% case, OmniRoute escalation for the 20% that needs frontier · cheapest total spend, fastest common path

Free forever (no token cap)

These providers don't publish a monthly cap — auto-fallback keeps them warm as the last-resort tier so auto always answers something.

Provider	Models	Terms	Auth	Best for	OmniRoute ID
1SiliconFlow siliconflow.com	Qwen, DeepSeek, GLM, Llama, Yi variants	Free tier, no cap, rate-limited	API key (free signup)	Broadest opensource selection on a single free provider	`siliconflow/qwen2.5-coder-32b`
2Z.AI (GLM-Flash) bigmodel.cn	GLM-4.5-Flash	Free forever · rate limited	API key	Fast general chat · Chinese/English	`zai/glm-4.5-flash`
3OpenCode Zen opencode.zen	Kimi K2, DeepSeek V3, Llama 3.3	Free · rate limited	None (keyless)	Zero-signup coding assistant · pre-wired in the `auto` combo	`oc/<model>`
4Kilo kilocode.ai	Sonnet-class + coding models	Free tier · no token cap	API key	Coding agent workloads	`kilo/<model>`
5Baidu Qianfan qianfan.cloud.baidu.com	ERNIE 4.5, DeepSeek, various	Free tier · rate limited	API key	Sovereign-friendly · Chinese-first	`baidu/ernie-4.5`
6Felo felo.ai	Search-augmented chat	Free · keyless	None (keyless)	Search-grounded answers · pre-wired in `auto`	`felo/<model>`

Free monthly quota

Documented monthly free tokens — reset on rollover. OmniRoute exhausts these first before touching paid tiers.

Provider	Free tokens / mo	Notable models	License / terms	Best for	OmniRoute ID
1Mistral La Plateforme console.mistral.ai	~1B / mo (Mistral Large 3)	Mistral Large 3, Codestral, Nemo, Mistral Small	Free tier · commercial-safe once paid	Highest single-provider free quota · European stack	`mistral/mistral-large-3`
2OpenAI (via free tools) platform.openai.com	~150M / mo (GPT-4o mini)	GPT-4o mini, o1-mini, embeddings	Free credits program	Frontier general-purpose · vision + reasoning	`openai/gpt-4o-mini`
3Google AI Studio aistudio.google.com	~60M / mo (Gemini 2.5 Flash)	Gemini 2.5 Flash / Pro, 1M-token context	Free tier · GCP terms	Long-context ingest · vision	`google/gemini-2.5-flash`
4Groq console.groq.com	~50M / mo · rate-limited	Llama 3.3 70B, DeepSeek R1 distill, Kimi K2, Qwen 2.5	Free tier · LPU inference	Sub-100ms latency · streaming	`groq/llama-3.3-70b-versatile`
5Cerebras cerebras.ai	~30M / mo · rate-limited	Llama 3.3 70B, Llama 4 Scout, Qwen 3 32B	Free tier · wafer-scale inference	Fastest tokens/sec on frontier open models	`cerebras/llama-3.3-70b`
6Cloudflare Workers AI workers-ai docs	Daily neurons · ~free tier	Llama, Mistral, Gemma, DeepSeek	Free tier · edge network	Edge-latency in 300+ cities · cheap per-token	`cloudflare/llama-3.3-70b`
7Anthropic (Claude Sonnet) console.anthropic.com	~25K / mo (Sonnet 4.5)	Claude Sonnet 4.5 / Haiku 4.5	Free trial credits	Deep reasoning · long context · frontier	`anthropic/claude-sonnet-4.5`

Signup credits (one-time)

First-month bumps that inflate the aggregate to ~2.15B tokens in month one. Non-repeating — flagged separately in OmniRoute's dashboard so they don't inflate the "steady-state" number.

Provider	Signup credit	Notable models	Terms	OmniRoute ID
Google Vertex AI	~300M / one-time	Gemini 2.5 Pro / Flash, PaLM	$300 GCP credit · 90 days	`vertex/gemini-2.5-pro`
AgentRouter	~200M / one-time	Aggregated frontier models	Signup credit	`agentrouter/<model>`
Predibase	~25M / one-time	Fine-tuned open models	Signup credit	`predibase/<model>`
Together AI	~25M / one-time	Llama 3.3, Kimi K2, DeepSeek, Mixtral, Qwen	$25 signup credit	`together/llama-3.3-70b`
Z.AI (GLM-CN)	~20M / one-time	GLM-4.5, GLM-Air, GLM-Flash	Chinese cloud tier signup	`zai-cn/glm-4.5`
Doubao · ByteDance	~15M / one-time	Doubao Pro, Doubao Lite	Signup credit	`doubao/<model>`
AI21 Labs	~10M / one-time	Jamba 1.5, Jurassic	Signup credit	`ai21/jamba-1.5-large`
LongCat	~10M / one-time	LongCat Flash, long-context	Signup credit	`longcat/<model>`
DeepSeek Platform	~5M / one-time	DeepSeek V3, R1	Signup credit	`deepseek/deepseek-chat`
Hyperbolic	~5M / one-time	Llama, Qwen, DeepSeek variants	Signup credit	`hyperbolic/<model>`
Nscale	~5M / one-time	Frontier open models	Signup credit	`nscale/<model>`

Cheap paid (when free runs out)

Tier 3 fallback. Pennies per million, but only touched when free tiers are exhausted and auto can't find a working free provider.

Provider	Price	Notable models	Terms	OmniRoute ID
1DeepSeek Platform platform.deepseek.com	$0.14 / M in · $0.28 / M out	DeepSeek V3, R1 reasoning	Pay-as-you-go · MIT-licensed weights	`deepseek/deepseek-chat`
2MiniMax minimax.chat	~$0.20 / M in · out	MiniMax-M1, MiniMax-Text-01	Pay-as-you-go	`minimax/<model>`
3Z.AI (GLM paid) bigmodel.cn	~$0.50 / M	GLM-4.5, GLM-4-Plus	Pay-as-you-go	`zai/glm-4.5`
4Kimi (Moonshot) platform.kimi.ai	$0.15 / M in · $2.50 / M out (K2)	Kimi K2, Kimi K3 (1M ctx)	Pay-as-you-go · Modified MIT weights	`kimi/k2`
5xAI (Grok) x.ai/api	$3 / M in · $15 / M out (grok-4)	Grok 4, Grok 3, Grok Imagine	Pay-as-you-go · closed weights	`xai/grok-4`

Bring-your-own-key (subscription tier)

Already paying for a plan? OmniRoute routes to it first, before any free tier. Your subscription quota gets used the way you already paid for it.

Plan / provider	What it gives	OmniRoute integration	OmniRoute ID
Claude Code plan	Anthropic Claude Sonnet / Opus via subscription	OAuth flow · uses your plan quota first	`anthropic-code/<model>`
Kimi Code plan	Kimi K2 / K3 via subscription	OAuth or API key	`kimi-code/k3`
GitHub Copilot Pro	GPT-4-class via Copilot subscription	Copilot API key	`copilot/<model>`
Codex plan	OpenAI Codex via plan	OAuth	`codex/<model>`
Any OpenAI-compatible API key	OpenRouter, Together, Fireworks, etc.	API key in OmniRoute dashboard	Provider-specific prefix

How to use these on vinOS.

OmniRoute is not shipped in the vinOS ISO (Node/npm dependency, moves fast). Install post-boot:

sudo vinos-install-aiEnsures Node + npm are present (from the ai bundle).

npm i -g omnirouteGlobal install. Server boots on localhost:20128.

omniroute startSystemd service starts on boot after this. Dashboard at http://localhost:20128.

export VINOS_AI_BASE_URL=http://localhost:20128/v1Point vinos-ai at the gateway. Add to ~/.bashrc to persist.

vinos-ai chat "hello"Answers via OmniRoute. Uses auto by default — routes to whatever's cheapest under quota right now.

omniroute statusLive view of provider health, quota remaining, active combos, saved tokens.

OmniRoute is upstream. vinOS just points at it.

We didn't build this gateway — diegosouzapw/OmniRoute did, and it's MIT-licensed. vinOS's contribution is the integration layer: vinos-ai honours VINOS_AI_BASE_URL, the routine TOML has an omniroute block, and a future vinos-install-omniroute bundle will one-shot the npm install + systemd unit + ~/.bashrc export. Support the upstream project on GitHub — they're the reason this page exists.

Download vinOS v2.1.0 Or run local via Ollama → OmniRoute on GitHub →